EX MCP Server

""" Explainable model selector for Auggie. Returns (model, reason) based on: - tool category - content hints - performance telemetry - user preferences/capabilities from auggie-config """ from __future__ import annotations from typing import Any, Optional, Tuple from src.providers.registry import ModelProviderRegistry as Registry from auggie.config import get_auggie_settings def _capability_score(name: str, caps: dict[str, Any], category: Optional[str], hints: dict[str, Any]) -> float: # Simple heuristic: category weight + hints speed = {"fast": 1.0, "medium": 0.6, "slow": 0.2}.get(str(caps.get("speed", "medium")).lower(), 0.6) reasoning = {"high": 1.0, "medium": 0.6, "low": 0.3}.get(str(caps.get("reasoning", "medium")).lower(), 0.6) cost = {"low": 1.0, "medium": 0.7, "high": 0.3}.get(str(caps.get("cost", "medium")).lower(), 0.7) score = 0.0 if category == "FAST_RESPONSE": score += 1.2 * speed + 0.5 * cost elif category == "EXTENDED_REASONING": score += 1.2 * reasoning + 0.4 * cost else: score += 0.9 * reasoning + 0.9 * speed if hints.get("code"): score += 0.1 # tiny boost for balanced models return score def select_model(category: Optional[str], hints: Optional[dict[str, Any]] = None) -> Tuple[Optional[str], str]: s = get_auggie_settings() or {} models = (s.get("models") or {}) available_cfg = models.get("available") or [] caps_cfg = models.get("capabilities") or {} # Discover from registry chain = Registry._auggie_fallback_chain(None) # If config lists available models, filter chain to those if available_cfg: chain = [m for m in chain if m in available_cfg] # Build scores explanations = [] scores = [] for m in chain: caps = caps_cfg.get(m, {}) sc = _capability_score(m, caps, category, hints or {}) scores.append((sc, m, caps)) if not scores: # fallback to preferred seed = Registry.get_preferred_fallback_model(None) return seed, f"Defaulted to registry preferred model: {seed}" scores.sort(reverse=True) best = scores[0] reason = f"Selected {best[1]} (score={best[0]:.2f}) based on caps={best[2]} and category={category}" # Incorporate simple telemetry nudge: prefer model with higher success rate if close tel = Registry.get_telemetry() or {} if len(scores) > 1 and tel: s0, m0, _ = scores[0] s1, m1, _ = scores[1] if abs(s0 - s1) < 0.15: def sr(m): t = tel.get(m, {}) succ = t.get("success", 0); fail = t.get("failure", 0) return (succ / max(1, succ + fail)) if sr(m1) > sr(m0) + 0.1: reason += f"; nudged toward {m1} by success-rate" return m1, reason return best[1], reason